vh
cd941f7bf9
Pre-clear snapshot. Rewrote Current state / in-flight to the post-rebalance topology (cyberprev is the sec seat, gen-small stood up, char/coder retuned, all generative seats native 262K in-band, catalog+bench shipped) and added a Recent decisions index entry + detail file 2026-09-14-fv-seat-rebalance-gen-small.md. Also commits two previously-untracked durable artifacts (fv-to-ana-nat runbook + its memory detail) so memory does not lag the work. Archival: file is over the ~300-line soft cap (377) but 0 entries archived — the 4 age-eligible (>14d) entries all carry open deferred-work pointers (guarded), and all other entries are <14 days old. Reported per the archival discipline.
4.3 KiB
4.3 KiB
2026-09-14 — fv-ml1 seat rebalance: cyberprev→sec, gen-small stood up, all seats to native 262K
Second big fv-ml1 session on 2026-09-14 (after the all-night reorg + orca blocker in
2026-09-14-fv-seat-reorg-and-orca-blocker). Everything here is committed; regenerate
the live view with scripts/seat-inventory.py and the durable catalog is
docs/pfi/llm-seat-catalog.md.
What changed (all committed)
- cyberprev quantized and promoted to the
secseat.hotdogs/Qwen3.8-27B-abliterated-cyber-preview(a cyber offensive-security tool-calling SFT finetune on an abliterated base — NOT "an abliteration"; trained onhotdogs/cyber-sft-agent-qwen38, 8,400 rows / 22 tools). ⚠ The raw hotdogs export ships broken: 850/1199 body tensors carry a tripledlanguage_model.prefix (unsloth export bug, verified in the upstream HF index) — the naive quant passed every structural check and served pure gibberish. Fix: name-repair the source against a known-good reference (exact 1199-tensor match), then house mixed-NVFP4 quant →cyberprev-nvfp4-mixed-v2. A coherence smoke is the gate; structural verification cannot substitute (see reference_unsloth_tripled_prefix_export_bug). Commits 1418edb/b94d173/196416f. - mog-sec RETIRED; cyberprev is now
sec/sec-reasoning. Gateway repointed (commit207798d).secis a ROLE alias, so this is a promotion, not silent substitution. ⚠ mog-sec was the fleet's ONLY offense+defense/blue-team seat — the sec role is now offense-only. mog-sec stack files retained for a one-command restore. Its "persona on stock" description was ALSO wrong — M.O.G.-SEC is itself an offense+defense SFT finetune (commit3906c68). All three sec candidates are Qwen3.8-27B finetunes differing by focus. - gen-small stood up (NEW seat).
llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-NVFP4-Experts-Only— already on disk at/tank/aimodels/qwen36-35b-a3b-heretic-nvfp4(no download). Heretic v1.3.0 (MPOA), 88% fewer refusals @ 0.0015 KL, 19 MTP. GPU0 :8026, MTP k=3 measured 69.6% accept / 3.09 len. 3.8 was requested but no general Qwen3.8 A3B exists → 3.6 is the fallback. Backssummarizer+classifier; newclassifier-large→ flash (accuracy tier);summarizer-largestays on flash. Commit05f3d3b. - All sec/gen/char seats raised to native 262,144 with concurrency in the 1.2-2.5x band (operator target). coder: 8192@13.75x → 16384@4.70x (waste reclaimed; ~4.7x overshoots 2-3x because the 1.5B weight floor sits under the util knob's resolution — a KV-byte pin would fix it, deferred). char-rp bumped 1.02→1.22x (util 0.40→0.52); GPU1 hit its ceiling so char-rp-fast's util cap was trimmed (its KV is pinned, so it held 2.04x) — an in-GPU rebalance rather than a utility-seat move.
- Durable docs:
docs/pfi/llm-seat-catalog.md(curated lineage/provenance/cards/speed)scripts/seat-bench.py(reproducible warm-tok/s + depth-OOM bench). Commitdfa91a8.
Measured — serial depth + warm tok/s (2026-09-14, warm/conc=1/greedy/n=3)
| seat | GPU | ctx | conc | warm tok/s | VRAM | depth clean to | OOM |
|---|---|---|---|---|---|---|---|
| cyberprev (sec) | 0 | 262144 | 1.37x | 136.6 | 47.1 GiB | 259,722 tok | none |
| gen-small | 0 | 262144 | 2.56x | 254.8 | 36.1 GiB | 254,526 tok | none |
| gen (flash-next) | 2 | 262144 | 1.31x | 170.7 | 95.3 GiB | 254,273 tok | none |
| char-rp | 1 | 262144 | 1.22x | 62.7 | 42.3 GiB | 254,858 tok | none |
| char-rp-fast | 1 | 262144 | 2.04x | 225.2 | 27.0 GiB | 254,823 tok | none |
| coder | 1 | 16384 | 4.70x | 337.3 | 5.6 GiB | 15,905 tok | none |
Every seat prefilled near max context with zero OOM/CUBLAS/illegal-memory (allocator-log
verified). Numbers are uncontended single-stream ceilings — re-derive with scripts/seat-bench.py.
Open / deferred
- mog-sec fate — retired; restore for blue-team coverage vs leave retired. Operator's call.
- coder ≤3x — needs a
--kv-cache-memorybyte pin (compose change) vs the fragile util floor. - gen-small 2.56x — a hair over the 2.5x band top; trimmable, left as useful headroom for the bulk seat.
- mog-sec / sentinel-r3 stacks retained on disk for restore; sentinel gateway aliases deleted (they 404).